Restructured for python utilities package, and made standalone utility to...

Restructured for python utilities package, and made standalone utility to generate labelled dataset that can easily be fed into keras later.

.gitignore

 build/
 src\openCV\CMakeLists.txt
 *\.venv
+dataset/
+sourceVideos/
\ No newline at end of file

src/openCV/main.cpp

@@ -20,7 +20,8 @@ ofstream createCSV(filesystem::path videoFilePath);
 //    and returns true for when the ratio of the sum of those changes is sufficiently large.
 bool detectChange(Mat firstFrame, Mat secFrame);
 
-double percentChangeThreshold = 0.00001;
+double percentChangeThreshold = 0.001;
+double avg = 0;
 
 int main(int argc, char** argv) {
   if (argc != 2) {
@@ -41,7 +42,7 @@ int main(int argc, char** argv) {
   VideoCapture decoder(videoPath.string());
   auto start = chrono::steady_clock::now();
 
-  int detectionsPerSec = 2;
+  int detectionsPerSec = 3;
 
   // Amount of extra time each detection interval will be padded with.
   double paddingTime_s = 1;
@@ -52,7 +53,9 @@ int main(int argc, char** argv) {
     maxPixelStability	maximum allowed credit for a pixel in history
     isParallel	determines if we're parallelizing the algorithm
   */
-  Ptr<BackgroundSubtractor> p_bgSub = bgsegm::createBackgroundSubtractorCNT(detectionsPerSec, true, 4*detectionsPerSec, true);
+  //Ptr<BackgroundSubtractor> p_bgSub = bgsegm::createBackgroundSubtractorCNT(detectionsPerSec, true, 4*detectionsPerSec, true);
+  Ptr<BackgroundSubtractor> p_bgSub = bgsegm::createBackgroundSubtractorGSOC( 0,20,0.003f,0.01f,32,0.01f,0.0022f,0.1f,0.1f,0.0004f,0.0008f );	
+
 
   // Pre-allocation of variables
   vector<bool> motionVec;
@@ -66,11 +69,14 @@ int main(int argc, char** argv) {
   Mat binaryNextFrame;
   Mat medianBlurFrame;    
   Mat medianBlurNextFrame;
-  int blurRadius = 9;
+  int blurRadius = 5;
 
   int32_t height = decoder.get(CAP_PROP_FRAME_HEIGHT);
   int32_t width  = decoder.get(CAP_PROP_FRAME_WIDTH);
 
+  int32_t newHeight = height/2;
+  int32_t newWidth = width/2;
+
   bool gotFrame, gotNextFrame;
   while (gotFrame = decoder.read(frame)) {
     // Skips some amount of "real" frames to reach desired "simulated fps".
@@ -82,6 +88,9 @@ int main(int argc, char** argv) {
     // Applies bg segmentation, median filter to reduce noise, and pushes the change detection result to output vector.
     // The frame count might be inaccurate at the moment, such that real frame 24 at time 1s, might be declared as being at time 0.9s. Needs testing.
     if (gotFrame && gotNextFrame) {
+      resize(frame, frame, Size(newWidth, newHeight), 0, 0, INTER_CUBIC);
+      resize(nextFrame, nextFrame, Size(newWidth, newHeight), 0, 0, INTER_CUBIC);
+
       p_bgSub->apply(frame, binaryFrame, -1);
       p_bgSub->apply(nextFrame, binaryNextFrame, -1);
 
@@ -94,11 +103,13 @@ int main(int argc, char** argv) {
   }
 
   decoder.release();
+  double test = avg / (double)simFrameTime;
+  cout << test << "\n";
 
   // Detection interval padding logic; essentially applies 1D dilation (morphological operation).
   int framesPaddedEachSide = ceil((paddingTime_s/2)*detectionsPerSec + 0.5);
   vector<bool> dilated = motionVec;
-  for(int i=0; i<motionVec.size(); i++) {
+  /*for(int i=0; i<motionVec.size(); i++) {
     if (motionVec.at(i) == 1) {
       for(int k = -framesPaddedEachSide; k <= framesPaddedEachSide; k++) {
         int index = i+k;
@@ -106,7 +117,7 @@ int main(int argc, char** argv) {
         dilated.at(index) = 1;
       }
     }
-  }
+  }*/
 
   // Timestamp generator logic; converts simulated frame index to real time in input file. As mentioned above, the math might be inaccurate.
   double timePerSimFrame_s = 1/(double)detectionsPerSec;
@@ -159,5 +170,7 @@ bool detectChange(Mat firstFrame, Mat secFrame) {
         sum += firstFrameArr[pos] | secFrameArr[pos];
     }
   }
-  return 4*sum/(firstFrame.rows*firstFrame.cols) > percentChangeThreshold;
+  double ratio = (double)sum/(255*firstFrame.rows*firstFrame.cols);
+  avg += ratio;
+  return ratio > percentChangeThreshold;
 }
\ No newline at end of file

src/python/RunDetection.py

@@ -14,7 +14,7 @@ def main(inputVideo):
         curDir = os.getcwd()
   
         #   Relative path to input folder from the current working directory (often the root of the repo).
-        repoRootToBinary = "build\\Release"
+        repoRootToBinary = ""
         pathToBinary = os.path.join(curDir, repoRootToBinary)
         os.chdir(pathToBinary)
 
@@ -27,7 +27,11 @@ def main(inputVideo):
             print("No input folder found. Creating it now. Place input video files in that folder.")
             return
         print("Processing file {file}".format(file=inputVideo))
-        os.system("opencvtest.exe {input}".format(input="input/{file}".format(file=inputVideo)))
+
+        #   Classical algorithm, C++ binary
+        #os.system("opencvtest.exe {input}".format(input="input/{file}".format(file=inputVideo)))
+
+
 
         prefix, ext = os.path.splitext(inputVideo)
         print("\nVideo processed and outputted to {csv}.".format(csv=prefix+".csv"))

src/python/VideoTrimmer.py

@@ -5,7 +5,7 @@ import logging
 import sys
 import ffmpeg
 import subprocess
-import Jaccard
+import Jaccard from utilities
 
 # Logger for logging purposes (seeing which file is currently being processed)
 logFormat = "%(levelname)s %(asctime)s - %(message)s"

src/python/groundTruth/groundTruth_Myggbukta-[2021-05-21_10-47-06]-037-gjedde_pa_19min.csv

+36,47
+192,199
+536,544
+604,616
+735,744
+873,887
+1141,1169
+1235,1244
+1296,1305
+1346,1354
+1453,1459
+1527,1532

src/python/groundTruth/groundTruth_Myggbukta-[2021-05-21_10-47-06]-399.csv

+44,51
+78,186
+191,202
+207,282
+284,294
+303,310
+318,337
+337,345
+346,357
+364,373
+380,388
+392,410
+410,429
+430,437
+439,457
+472,483
+493,534
+550,562
+565,568
+570,576
+580,625
+636,641
+650,681
+683,703
+720,753
+758,766
+825,836
+784,884
+890,927
+938,954
+957,978
+989,1023
+1042,1089
+1098,1114
+1121,1125
+1133,1144
+1145,1148
+1150,1156
+1160,1475
+1482,1685
+1687,1740

src/python/groundTruth/groundTruth_Myggbukta-[2021-05-21_10-47-06]-408.csv

+0,918
+929,934
+986,995
+997,1003
+1048,1056
+1056,1740

src/python/groundTruth/groundTruth_Myggbukta-[2021-05-21_10-47-06]-716.csv

+29,34
+34,86
+109,117
+126,132
+133,151
+152,448
+448,542
+574,585
+596,600
+612,621
+641,673
+698,727
+773,794
+903,908
+947,1047
+1142,1160
+1229,1251
+1295,1337
+1402,1416
+1420,1451
+1453,1469
+1469,1498
+1513,1543
+1629,1666
+1723,1740

src/python/utilities/GenerateDataset.py

+import cv2
+import csv
+import os
+import numpy as np
+import sys
+from pathlib import Path
+
+setFPS = 2
+resizeWidth = 480
+resizeHeight = 270
+
+groundsPath = "src/python/groundTruth/"
+
+###     Generates a dataset of labelled images from one video (soon a whole folder) by comparing to ground truth.
+###         Takes about 5 minutes for a 30 minute input.
+def generateDatasetFromFrames(inputVideoPath, desiredFPS):
+    decoder = cv2.VideoCapture(inputVideoPath)
+    inputVideo = Path(inputVideoPath).name
+
+    sourceNoExt = os.path.splitext(inputVideo)[0]
+    groundTruthName = "groundTruth_" + sourceNoExt + ".csv"
+   
+    # Looks for ground truth csv.
+    groundFound = False
+    for file in os.listdir(groundsPath):
+        if file == groundTruthName:
+            groundTruthCSV = open(groundsPath+file)
+            groundFound = True
+            break
+
+    if groundFound == False:
+        print("No ground truth file found to label from.")
+        return False
+
+    step = int(decoder.get(cv2.CAP_PROP_FPS) / desiredFPS)
+    try:
+        # Frame reading loop
+        while True:
+            ret, frame = decoder.read()
+            if not ret:
+                break
+
+            #   Frame index is one less because .read() gives index of next frame, not current.
+            #   Also does modulo step to simulate lower FPS.    
+            frameNo = int(decoder.get(cv2.CAP_PROP_POS_FRAMES)-1)
+            if frameNo % step == 0:
+                frame = cv2.resize(frame, (resizeWidth, resizeHeight), cv2.INTER_CUBIC)
+                curTime_s = int(decoder.get(cv2.CAP_PROP_POS_MSEC)/1000)
+
+                found = False
+                truthReader = csv.reader(groundTruthCSV)
+                for row in truthReader:
+                    #   Checks that the time, in seconds, of the current frame, is between an interval in the ground truth
+                    if (int(row[0]) <= curTime_s) and (int(row[1]) >= curTime_s):
+                        cv2.imwrite("dataset/fish/" + "frame{number}{source}.jpg".format(number=frameNo, source=sourceNoExt), frame)
+                        found = True
+                        break
+                
+                #   Resets csv iterator
+                groundTruthCSV.seek(0)
+                if found == False:
+                    cv2.imwrite("dataset/noFish/" + "frame{number}{source}.jpg".format(number=frameNo, source=sourceNoExt), frame)
+        print("Frames saved as labelled images in dataset folder.")
+        return True
+
+    finally:
+        decoder.release()
+        groundTruthCSV.close()  
+
+def main(video,fps):
+    print(generateDatasetFromFrames(video,fps))
+
+if __name__ == "__main__":
+    main("sourceVideos/Myggbukta-[2021-05-21_10-47-06]-408.mp4", setFPS) 
\ No newline at end of file

src/python/Jaccard.py → src/python/utilities/Jaccard.py

@@ -22,4 +22,6 @@ def calculateSuccess(observedIntervalsList, trueIntervalsList):
 
     union = cumulativeObs + cumulativeTrue
     index = cumulativeIntersection / union
+    ratioObsToReal = cumulativeObs / cumulativeTrue
     print("    Correctness: {:.4f}".format(index) + ", or {:.1f}%".format(index*100))
+    print("    Observed-to-expected ratio: {:.2f}".format(ratioObsToReal) + " --> {sensitivity}".format(sensitivity=("Too sensitive" if ratioObsToReal>1 else "Not sensitive enough")))
\ No newline at end of file